Semantic segmentation of deep-sea polymetallic nodules based on an improved DeepLabV3+ algorithm

Due to the complexity of the deep-sea environment and the diversity of target features, traditional image segmentation models face challenges in deep-sea operations. To address the limitations in feature extraction, segmentation accuracy, and model lightweighting, this study proposes an enhanced DeepLabV3+-based semantic segmentation model. The model uses MobileNetV2 as the lightweight backbone, significantly reducing the parameter count. A new convolutional layer in the ASPP module with optimized dilation rates improves multi-scale feature representation. The model also integrates PSA (Point-wise Spatial Attention) and CBAM (Convolutional Block Attention Module) to enhance the perception of both global contextual information and local details. Experiments were conducted on high-resolution deep-sea imagery captured by the "Jiaolong" manned submersible. The proposed model achieves an Intersection over Union (mIoU) of 84.1%, Pixel Accuracy (mPA) of 91.44%, and Precision (mPrecision) of 91.21%. With a compact size of just 24.2 MB, it balances segmentation performance and lightweight design more effectively than mainstream models. The results demonstrate that the proposed model offers superior segmentation accuracy, robustness, and resource efficiency, making it a practical solution for deep-sea polymetallic nodule segmentation.